Thread by @EliotHiggins

Making pictures of Trump getting arrested while waiting for Trump's arrest.

This blew up etc, so donate to Bellingcat so we can catch bad guys and make scenes like this real bellingcat.com/donate/

Meanwhile, in AI Trump world:

Apparently doing a viral thread about Donald Trump using Midjourney images is enough to get you banned from Midjourney, oops.

At least I still have ChatGPT 4, so here's the Trumpshank Redemption for those of you who have stuck around this long.

One step closer to Trump seeing this thread.

Now I'm banned from Midjourney you'll be spared my Boris Johnson noir detective story

A few things I noticed when generating these images. Midjourney is good at understanding simple actions, but slightly more complex and obscure actions get odd results. This is "Donald Trump wearing an orange prison jumpsuit carving a key out of soap"

Something as simple as "Donald Trump getting peppersprayed" seemed to confuse it a lot. I get the feeling it's drawing on images of people pouring milk in their eyes to counteract the effects of riot control agents.

However, if you use simple actions and common words you can get good results, like "Donald Trump in an orange prison jumpsuit running away from a burning mansion at night in the rain"

I spoke to the Washington Post about all this nonsense washingtonpost.com/politics/2023/…

Another point about these images is despite some people thinking this is the end of being able to spot fake photos, is that Midjourney just can't recreate real locations accurately, like the European Court of Human rights in this example. The real one is on the right.

When we teach open-source investigations in workshops geolocation is one of the core skills we teach, so if you can't geolocate a photo then it loses a lot of its value (if not all value) as evidence, so in terms of evidence Midjourney images aren't a real problem.

Creating something like a realistic McDonald's restaurant was very difficult, mainly because Midjourney can't handle text very well in images. Here I had to specifically ask for the golden arches to be included, which likely rendered successfully because it's such a distinct… twitter.com/i/web/status/1…

However, there's a difference between images as evidence and images as information. A fake image (with or without AI's involvement) can go viral far more quickly than they can be fact checked in most cases.

Fake videos were quickly addressed at the start of Russia's 2022 invasion of Ukraine because of ad hoc communities who came together online who rapidly debunked them, but that's only because of long term engagement of people interested in Ukraine in topics related to… twitter.com/i/web/status/1…

So the solution is more education around critical thinking and raising awareness about how to spot fake images. Even with that you'll always have people who share fake images uncritically because it reinforces their own beliefs.

Does this mean Trump will get banned from Midjourney too?

https://twitter.com/paleofuture/status/1638969629409431552

A very good take on this thread and its implications.
theatlantic.com/technology/arc…

Good piece from the Washington Post on Midjourney and AI imagery
washingtonpost.com/technology/202…

Poetic.